AI4EO: from big to small architecture for deployment at the edge

This presentation aims to provide an overview of the required preparatory steps for an effective deployment of AI at the edge for Earth Observation mission. It is based primarily on the lessons learnt from Phisat-1, an enhancement of the Federated Satellite Systems (FSSCat) mission. Launched in September 2020, it was the first experiment to demonstrate how AI can be used for Earth observation. More specifically, an AI inference engine for cloud detection has been demonstrated thanks to an Intel® Movidius™ Myriad™ 2 Vision Processing Unit (VPU). In this presentation, we will also introduce the capabilities of the planned Phisat-2 mission (with Multispectral EO sensor with 7-bands in VNIR range).

More specifically, we will also address the following topics in this presentation:
- An analysis of the need for AI@edge is provided starting from end–user requirements. The AI methodologies (for instance mainly Deep Neural Network) being mostly supervised techniques, relevant databases for training are required. However, self-supervised or weakly supervised AI are of interest when sparse or no annotation / ground truth are available.
- We will also provide an overview of AI techniques to cope with uncertainties, especially when it is required to use synthetize EO data (e.g. multispectral optical image) that will be acquired by the upcoming flying platform subject to noise. New approaches can be investigated (e.g. NN Out-of-distribution detection-OOD, Bayesian DL) with the general idea to provide robust estimates with quantified uncertainties. Bayesian DL refers to merging deep learning architectures and Bayesian probability theory. Bayesian DL models typically derive estimations of uncertainty by either placing probability distributions over model weights, or by learning a direct mapping to probabilistic outputs.
- The design of Tiny ML models via binary dropout, pruning, and knowledge distillation is of high interest so as to be compliant with on-board requirement.
- We will report on current activities about the design of neuromorphic algorithms (spiking neural networks) for solving EO problems such as Temporal coding or rate-based coding)